chq2018515 / learn-to-defer

Code for "Consistent Estimators for Learning to Defer to an Expert" (ICML 2020)

Consistent Estimators for Learning to Defer to an Expert

This repository includes the code and experiments for our ICML 2020 paper Consistent Estimators for Learning to Defer to an Expert by Hussein Mozannar and David Sontag.

Quick Links:

• CIFAR
• Hate speech detection
• Synthetic Data
• CheXpert

This repository is currently being expanded.

Learning to defer to an Expert

Learning algorithms are often used in conjunction with expert decision makers in practical scenarios, however this fact is largely ignored when designing these algorithms. In this project we explore how to learn predictors that can either predict or choose to defer the decision to a downstream expert. Our approach is based on a novel reduction to cost sensitive learning where we give a consistent surrogate loss for cost sensitive learning that generalizes the cross entropy loss. The loss described in detail in the papers is referred to as L_{CE}^{\alpha}. This repository contains multiple jupyter notebooks

Requirements

We include a requirements file that covers everything required to run the notebooks from a new environment.

The major dependencies are the following:

jupyterlab
torch
torchvision
torchtext
scipy
matplotlib
scikit-learn

Citation

@inproceedings{mozannar2020consistent,
    title={Consistent Estimators for Learning to Defer to an Expert},
    author={Hussein Mozannar and David Sontag},
    booktitle = {Proceedings of the Thirty-Seventh International Conference on Machine Learning (ICML)},
    year={2020},
    eprint={2006.01862},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url_Paper = {https://arxiv.org/pdf/2006.01862.pdf}
}